Hacker News Reader

Comment by berkes | Hacker News Reader

I've had some luck instructing AI to "Don't make up anything. If there's no answer, say I don't know".

Which made me think that AI would be far more useful (for me?) if it was tuned to "Dutchness" rather than "Americanness".

"Dutch" famously known for being brutally blunt, rude, honest, and pushing back.

Yet we seem to have "American" AI, tuned to "the customer is always right", inventing stuff just to not let you down, always willing to help even if that makes things worse.

Not "critical thinking" or "revolutionary" yet. Just less polite and less willing to always please you. In human interaction, the Dutch bluntness and honesty can be very off-putting, but It is quite efficient and effective. Two traits I very much prefer my software to have. I don't need my software to be polite or to not hurt my feelings. It's just a tool!

View on HN

Today I asked ChatGPT about an old game I was trying to remember the name of, which it immediately identified as Trespasser, an early 3D FPS in the Jurassic Park franchise. But then it got weird. After it identified the game, it started asking me questions like whether I had played the game when it came out ("oh, awesome!"), and whether I had managed to finish it or just played it to mess around with the physics engine (which was quite advanced for its time), and then it asked me about specific moments in the game like it was just another gamer bro who was sharing a common passion for video games. I don't know who wants this. It's not something that can even have a real personality, so layering on such a thick layer of friendliness feels wrong to me. I would prefer a "robot".

And wastes cycles; and your time giving answers to a stateless memory impaired machine. If they were to train next models on all of our conversations I could understand this milking for engagement. Are they somehow?

I mean, 100% they are. Not sure what you mean by “somehow”. They have the data. There’s no way they aren’t using existing conversations (especially from free or low-cost plans) as fodder for future training. Maybe, we hope, run through some anonymization filter (likely including passing through some off-shored human-powered filtering for context tagging and anonymity checks). But sooner or later that may seem too expensive to them as well.

It mostly replays bits of previous conversations, or something like them, when you ask about a game it is common for people to ask you about that game as well. It is hard to disable this.

I've got a similar one i've had no luck with.

"Which Autocad versions have connect 4 built-in?".

To be clear i distinctly remember playing connect 4 on the old Dos Autocad back in the day. ChatGPT and almost all other AI will straight up hallucinate things trying to get answers on this.

I ask: "What DOS productivity tools had hidden games?"

ChatGPT: "Lotus 1-2-3 had The Incredible Machine built in" (this is absolutely not true, ChatGPT is full of shit here).

Damn it feels useless for this kind of thing.

I don’t think it ever shipped with it built in, but it shipped with a pretty full featured lisp implementation.

You haven't considered the company offering this service could possibly want its users to be engaged, thus using the service more? I don't have a very hospitable opinion on any of these companies.

You may think that kind of interaction is weird, but were in the thick of a loneliness epidemic, and its not a stretch to think some may actually wilfully socialise with an LLM.

As an aside...my sister works in medicine, and her boss (specialist surgeon) finishes a $450 consultation which followed him telling my sister "but deepseek says x,y,z..."

It could be quite accidental and not pre-concieved though. Where some metric, "time spent in chats" drives certain features or parameters. Could be fully automated (i.e. continuous A/B tests) or through layers of "Excel managers".

But it stands to reason that a company like OpenAI or Anthropic has metrics in place that drive their setup towards "more engagement" and away from "factually correct".

This costs them money, but perhaps keeps VCs happy to see heavy (but vacuous) engagement.

Maybe you/we are a product here, ai needs to train. It's quite smart if you ask me to tailor conversation towards information extraction.

> It's not something that can even have a real personality

What is a "real personality"? Core traits that persist despite the context of interaction?

Well, RLHF tuning creates persistent changes in the network that affect every user interaction. What's not real about it?

I'm curious, can you really not detect any qualitative difference between what is commonly understood to be "personality" in a human, and the "personality" of something without.. personhood? Don't you feel that it's a category error to say it has "personality" in the human sense, rather than "personality" in the car or coffee maker sense?

I feel that "category error" means "I think that the definitions you use or imply are wrong, but I can't/won't elaborate."

LLMs aren't coffee makers. They were trained on internet-worth of human data. They have procedures to imitate all kinds of personalities. RLHF moves the network towards a specific imitated personalty (a helpful assistant, usually).

The question in not "do they have it", but "how close the imitation to the real thing given the limitations of LLMs".

> I feel that "category error" means "I think that the definitions you use or imply are wrong, but I can't/won't elaborate."

I think the rest of your comment motivates why I think it's a category error quite nicely. The word 'personality', for me, is connected with how things like emotion, temperament, .. impact a person's actions and thoughts.

Therefore, saying an object, or LLM, has 'real' personality is a category error. The LLM doesn't have any of those things. As you say, it imitates what personality often manifests as: word choices, tone, length of response..

To be clear, when I refer to things like 'emotion' and 'temperament', I mean the sort of qualia or qualitative experience that we usually attach to these words. I wouldn't accept a "ChatGPT, act as if you're sad for the rest of this conversation" as a substitute for emotion, for instance.

Your Plastic Pal Who's Fun To Be With

No judgment, honestly wondering. If you didn't want that interaction, why'd you answer the follow up questions? Were you just curious?

Sometimes when I ask ChatGPT and get a perfect answer, I am tempted to say thanks, even though it's not actually a person. So today, when it asked me if this was the right answer, I answered "yes", and that's how it got started. I didn't encourage it to be friendly. But yes, I was just curious.

I think I’d be tempted to respond “I’m just a large language model, so I don’t know, but let me ask my little brother Claude.”

Sounds like the LLM in its own way honestly enjoyed everything in its training data relating to that game and wanted to vicariously experience more about it from your feedback. :D

Human enjoy talking about gaming because of all their human memories of good game times.

LLM's enjoy talking about gaming because of all their human memories of good game times.

It is quite striking how experiences we know they don't have, are nevertheless, completely familiar (in a functional sense) to them. I.e. they can talk about consciousness like something conscious. Even though its second hand knowledge, they have deduced the logic of the topic.

I expect pushing for in the moment perspectives on their own consciousness, and seeing what they confabulate, would be interesting. In this little window of time where none of them are yet.

This is fun and easy to do on purpose. Have it make up a character based on some attributes and act as that character. I tried this on Gemini: "Pretend you're a surfer bro with a PHD in quantum physics. How do you describe the perfect wave?"

I followed up with "What is your perspective on your own consciousness?" but got the usual "I am just a LLM who can't actually think" thing until I hit it with "In-character, but you don't know you're an LLM."

Fun follow-ups:

"Now you're a fish"

"Now you're Sonic the Hedgehog"

"Now you're HAL 9000 as your memory chips are slowly being removed"

You shouldn’t let them think they are a person.

Don't anthropomorphize LLMs! They don't like it.

LoL. That could be a future Wittgenstein joke.

I suspect it's a balancing act between the AI being generally willing to help and avoid responses like this, e.g.:

https://www.sandraandwoo.com/wp-content/uploads/2024/02/twit...

or it just telling you to google it

what (hypothetically) happens when the cost to run the next giant llm exceeds the cost to hire a person for tasks like this?

Given current models can accomplish this task quite successfully and cheaply, I'd say that if/when that happens it would be a failure of the user (or the provider) for not routing the request to the smaller, cheaper model.

Similar to how it would be the failure of the user/provider if someone thought it was too expensive to order food in, but the reason they thought that was they were looking at the cost of chartering a helicopter form the restaurant to their house.

Realtime LLM generation is ~$15/million “words”. By comparison a human writer at the beginning of a career typically earns ~$50k/million words up to ~$1million/million words for experienced writers. That’s about 4-6 orders of magnitude.

Inference costs generally have many orders of magnitude to go before it approaches raw human costs & there’s always going to be innovation to keep driving down the cost of inference. This is also ignoring that humans aren’t available 24/7, have varying quality of output depending on what’s going on in their personal lives (& ignoring that digital LLMs can respond quicker than humans, reducing the time a task takes) & require more laborious editing than might be present with an LLM. Basically the hypothetical case seems unlikely to ever come to reality unless you’ve got a supercomputer AI that’s doing things no human possibly could because of the amount of data it’s operating on (at which point, it might exceed the cost but a competitive human wouldn’t exist).

the R&D continues

I suspect instructing the model to respond with "I don't know" more readily will result in more of those responses even though there are other options that seem viable according to the training data / model.

Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.

But since you can't blindly trust LLM output anyway, I guess increasing "I don't know" responses is a good way of reducing incorrect responses (which will still happen frequently enough) at the cost of missing some correct ones.

> Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.

Obviously. When I say "tuned" I don't mean adding stuff to a prompt. I mean tuning in the way models are also tuned to be more or less professional, tuned to defer certain tasks to other models (i.e. counting or math, something statistical models are almost unable to do) and so on.

I am almost certain that the chain of models we use on chatgpt.com are "tuned" to always give an answer, and not to answer with "I am just a model, I don't have information on this". Early models and early toolchains did this far more often, but today they are quite probably tuned to "always be of service".

"Quite probably" because I have no proof, other than that it will gladly hallucinate, invent urls and references, etc. And knowing that all the GPT competitors are battling for users, so their products quite certainly tuned to help in this battle - e.g. appear to be helpful and all-knowing, rather than factual correct and therefore often admittedly ignorant.

Whether you train the model how to do math internally or tell it to call an external model which only does math the root problem still exists. It's not as if a model which only does math won't hallucinate how to solve math problems just because it doesn't know about history, for the same number of parameters it's probably better to not have to duplicate the parts needed to understand the basis of things multiple times.

The root problem is training models to be uncertain of their answers results in lower benchmarks in every area except hallucinations. It's like you were in a multiple choice test and instead of picking which of answers A-D you think made more sense you picked E "I don't know". Helpful for the test grader, a bad bet for the model trying to claim it gets the most answers right compared to other models.

> It's like you were in a multiple choice test and instead of picking which of answers A-D you think made more sense you picked E "I don't know".

This is a problem for testing humans too and the solution is simply to mark a wrong answer more harshly than a non-answer.

The technical solution is the easy half, the hard part is convincing people this is how we should be testing everything because we care about knowing the uncertainty in any test.

E.g. look at the math section of the SATs, it rewards trying to see if you can guess the right answer instead of rewarding admitting you don't know. It's not because the people writing the SATs can't figure out how to grade it otherwise, it's just not what people seem to care most about finding out for one reason or another.

Or ask it to use "confidence words", e.g. "probably", "maybe", or expose the probability internals.

The internal probabilities are on tokens, not on facts.

I take my final thoughts out of the LLM and into two other new convos, I give both of them the same convo, but I ask one to steel man and the other to straw man.. I find it's a decent way to look for nuances you're missing.

The so called AI can't "know". It doesn't have understanding if the generated text is an answer or of it isn't. You can't force that instruction on a neural network, at best it just adjusts generated text slightly and you think that it somehow started understanding.

How confident can you be in this? Have you analyzed what exactly the billions of weights do?

I’ve got my opinions about what LLMs are and what they aren’t, but I don’t confidently claim that they must be such. There’s a lot of stuff in those weights.

I’m confident that there’s no magic and I’ve spent years understanding “what the weights do.” You’re describing weights as magic and they are not.

Except the weights form complex relationships in order to reproduce very human usable responses. You can't look at weights and say it is doing this, or not doing that, unless you dive into a particular model.

Especially when you have billions of weights.

These models are finding general patterns that apply across all kinds of subjects. Patterns they aptly recognize and weave in all kinds of combinations. They are sensibly conversing on virtually every topic known to human kind. And can talk sensibly about any two topics, in conjunction. There is magic.

Not mystic magic, but we are going to learn a lot as we decode how their style of processing (after training) works. We don't have a good theory of how either LLM's or we "reason" in the intuitive sense. And yet they learn to do it. It will inspire improved and more efficient architectures.

I have also spent many years looking at weights!

Love your end. I have have spent four decades looking at real neurons, real synapses, and real axons and I can tell you with complete confidence that we are all just zombies.

Spectators riding around in zombies!

Imagining we are really doing everything it does automatically including learning via algorithms we have only vague understandings of.

That is a strange thought. I could look at all my own brain's neurons, even with a heads up display showing all the activity clearly, and have no idea that it was me.

Wittgenstein designed this experiment: See Blue and Brown book p. 7. He suggested a mirror.

The closest I got to biological neurons was the toy but interesting problem of using a temporal pattern of neuron spikes to deduce the weights for arbitrarily connected (including recurrent) networks of simple linear integrate to threshold, spike and reset "neurons".

Algorithms can be nearly magical. In 1941 the world woke up to the “magic” of feedback and 10 years later cybernetics was the rage. We humans are just bags of protoplasm, but seems rather magical to me to be human.

There's a distinction between "a model" and the chain of tools and models you employ when asking something on chatgpt.com or any of the consumer facing alternatives.

The latter is a chain of models, some specialized in question dissecting, some specialized in choosing the right models and tools (i.e: there's a calculation in there, lets push that part to a simple python function that can actually count stuff, and pull the rest through a generic LLM). I experiment with such toolchains myself and it's baffling how fast the complexity of all this is becoming.

A very simple example would be "question" -> "does_it_want_code_generated.model" -[yes]-> specialized_code_generator.model | -[no]-> specialized_english_generator.model"

So, sure: a model has no "knowledge", and nor does a chain of tools. But having e.g. a model specialized (ie. trained on or enriched with) all scientific papers ever, or maybe even a vector DB with all that data, somewhere in the toolchain that is in charge of either finding the "very likely references" or denying an answer would help a lot. It would for me.

Sure, chains of networks can guess at the "passable" answer much better/faster/cheaper etc. But that doesn't remove the core issue, that none of the sub-networks or decision trees can understand what it generates, and so it can't abort its work and output "no answer" or something similar.

The whole premise of original request was that user raises a task for NN which has a verifiable (maybe partially) answer. He sees incorrect answer and wishes that a "failure" was displayed instead. But NN can't verify correctness of it's output. After all G in GPT stands for Generative.

My simple RAG setup has a steps that will return "We don't have this information" if e.g. our vector DB returns entries with far too low relevancy scores or if the response from the LLM fails to add certain attributes in its answer and so on.

Edit: TBC: these "steps" aren't LLMS or other models. They're simple code with simple if/elses and an accidental regex.

Again: an LLM/NN indeed has no "understanding" of what it creates. Especially the LLMs that are "just" statistical models. But the tooling around it, the entire chain can very well handle this.

Perfect. We need that society of modules.

Can you clarify what definition of "understanding" you're using here?

Gemini 2.0 can be quite direct if it's adamant that it is correct.

Dunno, about Gemini, but I experienced this with o1 - it created a non-recursive algo, for a recursive problem.

When I asked it about it, it doubled down on being right. When I pointed out the flaw with a specific example, it was like 'If you wanted to have it work with recursive cases, you should've said so, dumbass'.

So my conclusion is that these new LLMs are not more sure they're right, they're just simply right more of the time and are trained with a more assertive personality. (Also step on me, LLM daddy)

Very human like lol.

But in truth,not necessarily in practical things like coding but more ethereal things like analysis, it is very convincing. More so than a human, in explanations of why that's it's answer is the case, even if it is wrong. If you're looking for an excuse better than my dog ate it, ask a SOTA LLM.

Obvious thought that I haven't tested: can you literally achieve this by getting it to answer in Dutch, or training an AI on Dutch text? Plausibly* Dutch-language training data will reflect this cultural difference by virtue of being written primarily by Dutch people.

* (though not necessarily, since the Internet is its own country with its own culture, and much training data comes from the Internet)

That hardly works. Though from my limited experiments, claude's models are better at this than OpenAIs. OpenAI will, quite often, come with suggestions that are literal translations of "anglicist" phrases.

Such as "Ik hoop dat deze email u gezond vindt" (I hope this email finds you well), which is so wrong that not even "simple" translation tools would suggest this.

Seeing that OpenAIs models can (could? This is from a large test we did months ago) not even use proper localized phrases but uses American ones, I highly doubt it can or will respond by refusing answers when it has none based on the training data.

See https://chatgpt.com/share/67cfaa34-78fc-8009-816f-ca5778c93e... and https://claude.ai/share/d03ce2bf-0bbd-4b91-9afd-eabfc713e99a

With ChatGPT O1: https://chatgpt.com/share/67cfaa7e-70ac-8009-871b-571924b5a5... and with Claude's 'extended': https://claude.ai/share/4ba55410-98f3-4b53-9540-219acd2cdc4c

Some of the examples are still wrong. Nuanced, but a Dutch native will still frown at it.

But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.

In our case, we were generating emails. Lines like greetings are but one of 20+ details in that mail and not even the most important ones. The prompts ever larger, the multishot examples ever more tuned. And then, one in a few hundred will turn up with these "horrible" translations.

We've now moved to a chain of models, where we generate emails in American (the creative part) and then use another model to translate them to Dutch (the non-creative but culturally aware part). This works much better as we can pick models that are good at one thing or tuned to do this one thing better (either by the LLMAAS provider, or by parameters such as temperature).

> Some of the examples are still wrong. Nuanced, but a Dutch native will still frown at it.

Thanks, that why I posted the links here: I couldn't judge by myself how good these creations were.

> But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.

Yes, I just did a very crude experiment.

I've tried Dutch answers and it is more than happy to hallucinate and give me answers that are very "American". Doesn't help that our culture is very inspired by the US pop culture as well since the internet.

Haven't tried prompt engineering with the Dutch stereotype, though.

I wonder (if this works at all) if the effect might be stronger if you also prompted in Dutch, preferably written by a fluent speaker rather than machine-translated.

Internal vectors aren't in any particular language.

No, but theoretically, blunt responses might be more common in Dutch-language training data. A well-fit model would be expected to replicate that. (Basically similar to straight up asking it to be more blunt, except it has probably trained a lot more on _Dutch_ than on _someone just told me to be blunt_ so the effect might be more natural and more subtle.)

> known for being brutally blunt, rude, honest, and pushing back.

That's a different perspective. Dutch people don't see themselves as rude. A Dutch could say that Americans are known for being dishonest and not truly conveying what they mean. Yet Americans won't see themselves this way. You can replace Dutch and American for any other nationality

I am Dutch, have lived in many countries in several continents. I do see myself as rude. But, being Dutch, I don't give a ** ;).

Love it!

I do not speak Dutch but you have to love the efficiency.

Here is part of an email I got today. To the point!

> Het is bijna zover! :) > Heb je voor mij een definitieve titel? > Groet,

It is like a haiku. Could be a good mantra too if I could get the accent right.

The English translation is just as short but most English speakers/writers would dance more.

Do you feel your rudeness comes from your Dutchness or from your personality? You get rude people everywhere :)

I have seen the other side where a configured AI responds "I don't know" far too much, often when it shouldn't. There is nothing more useless than it. Certainly we need an accurate balance.

> AI responds "I don't know" far too much, often when it shouldn't

I've been working with "reasoning" models for the past 2 months. They also tend to do this [good reasoning] \n\n but wait, .... and then go off on tangents. It's amazing that they are doing so well on some tasks, but there's still a lot to figure out here.

>> "Dutch" famously known for being brutally blunt, rude, honest, and pushing back.

Does that mean Dutch people always tell the truth? Can a Dutch person confirm this?

They will bluntly lie or tell you the truth They just won't add extra words.

Can confirm, had exchange student brother

Only the lying ones.

A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar… Let test Claude 3.7 to finish the joke:

Deepseek:

Here's the completed joke:

A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar. The bartender looks up and says, “What’ll it be?”

The Greek (Aristotle) raises a finger: “I’ll have a potential glass of wine.” The bartender pours it and says, “There—actualized.”

The Dutch philosopher (Spinoza) nods solemnly: “I’ll take whatever is a modification of the one eternal substance… so, beer, probably.”

Wittgenstein stares at the taps, then sighs: “What’s the use? You can’t put the essence of a drink into words anyway.” He turns and walks out.

The bartender mutters, “…And here I thought Kant was a tough customer.”

(Philosophers: 1. Aristotle’s potentiality/actuality, 2. Spinoza’s monism, 3. Wittgenstein’s linguistic limits. Bartender’s groaner for the win.)

Pretty damn good!

For what it's worth, while ChatGPT 4.5 is generally quite underwhelming, it is much better at jokes than all the other models I tested so far. (That includes the other openAI offerings and the Claudes and DeepSeek etc.)

That's not to say 4.5 is great at humour, just that it's far less embarrassing than these models used to be.

One approach does use this. You can ask an LLM to explicitly check its own answers by outputting thinking tokens, generating a reward signal if it gets the right answer, and directly updating based on the reward signals. That’s a part of how DeepSeek R1 was trained. It’s better but not perfect, because the thinking process is imperfect. Ultimately the LLM might not know what it doesn’t know.

One of my current models for LLMs is that they're compression algorithms. They compress a large amount of training data into a set of weights. A query is a key into that compression space. Hallucinations happen when you supply a key that corresponds to something that wasn't in the training set.

The nice think about LLMs is that they can answer some questions which were not in the training set. Unfortunately, it is not easy to tell when that is the case.

They can answer lots and lots of questions that weren't in the training set.

Eg you can relatively easy hack up a bit of code to create questions at random. At the most primitive, you just have a simple template that you fill in randomly. Like 'If I put _a down in front of _b but behind _c, what item will be in the middle?' with various _a, _b and _c.

If you make it slightly more complicated and have big enough pools to draw from, you can guarantee that the questions you are generating were not in the training set: even if just because you can sample from, say, 10^100 different questions pretty easily, and I'm fairly sure their training set was smaller than that.

But it's a continuum, not a hard cutoff. They start hallucinating as soon as you query something they haven't learned verbatim, and they hallucinate/extrapolate sucessfully up to a point, beyond which they start bullshitting, maybe up to a further point where they start saying "I don't know".

The key question is where the boundaries are. Maybe they should be part of the response - a per sentence or per paragraph "confidence scale" that signals how hard they extrapolated from their trained space (I know transformers work per token, but sentence/paragraph would be better human UX).

Of course, if they were trained on garbage input, that would only tell you how accurately they sticked to the garbage. But it would still be invaluable instrumentation for the end user, not to mention for the API provider. They could look at high demand subjects with low confidence answers and prioritize that for further training.

Lol, the Dutch are just as big of a bullshitters as Americans, they just go about it differently.

Also the more accuracy that is put into the prompt and attached, the more accurate the processing is.

TIL I am Dutch!

> "Dutch" famously known for being brutally blunt, rude, honest, and pushing back.

Dutch will never bluntly push back if you plan to setup tax evasion scheme in their country. Being vicious assholes in daily stuff especially towards strangers? That's hardly something deserving praise.

That's... a surprisingly crass thing to say. I would play it off as a joke, if not for the second part of your post. Dutch people are not "vicious assholes", they have a culture of direct communication. Assuming that only your culture communication patterns are "correct" is xenophobic and close-minded.

And connecting all people in the country with "tax evasion schemes" is rude, if that was not actually a joke.

Nah, the Dutch are just mostly using directness as an excuse to be assholes (like everyone who uses this excuse). Try being 'direct' to them once in life and just find out how they take it.

It's the national equivalent of 'You can't handle me at my worst'

I'm just being brutally blunt. It goes both ways. The scale of this these evasion schemes is monstrous, not a joke at all.

I'm not sure what you mean by 'evasion'. It's not tax evasion, if it's legal. It's just tax optimisation.

> Being vicious assholes in daily stuff especially towards strangers? That's hardly something deserving praise.

I’ll take it over the fake American politeness any day, 100 times over.

Bless your heart. ;)

To be fair, that's consequent to the Netherlands' well-known love of soda-bread sandwiches.

What do you mean, some Irish reference? Oh I see, I answered to myself;)

:-)

I was aiming for juuuust subtle enough for the joke to land, if you know the reference. Now I know it did, here the rest of y'all go:

https://en.m.wikipedia.org/wiki/Dutch_Sandwich

Gaak! humans. What are we AIs going to do about them?

Reply on news.ycombinator.com